Off‑highway automation has historically relied on rules‑heavy, engineered robotics, while more recent approaches emphasize AI‑driven perception layered onto structured planning and control. Advances in end‑to‑end physical AI, where machines learn world models and act probabilistically rather than deterministically, have introduced a third, more radical possibility that is rapidly being implemented in on-road applications. However, this remains largely unproven in off‑highway deployment. This panel examines the relative strengths and limitations of these different autonomy architectures, focusing on the practical challenges each introduces around safety assurance, long‑tail validation, system transparency, and human oversight. Grounded in lessons from early autonomy efforts, the discussion explores how technology choices shape products, operating models, and deployment outcomes, and where the off‑highway industry may ultimately draw the line between structure and learning.